Depth Map Recovery Based on a Unified Depth Boundary Distortion Model.

Area of Science:

• Computer Vision
• Image Processing
• 3D Reconstruction

Background:

• Depth maps from sensors or learning methods often suffer from boundary distortions.
• These distortions include missing, fake, or misaligned boundaries compared to RGB images.
• Accurate depth boundary representation is crucial for 3D scene understanding.

Purpose of the Study:

• To propose an RGB-guided depth map recovery method.
• To address and correct serious boundary distortions in depth maps.
• To improve the accuracy and alignment of depth boundaries with RGB images.

Main Methods:

• A unified model was developed to identify erroneous regions in distorted depth maps.
• Local structures from RGB and depth maps were extracted using Gaussian kernels and compared via SSIM index.
• A depth map recovery method iteratively identifies and corrects erroneous regions using the unified model and a weighted median filter.
• Texture-copy artifacts were mitigated by focusing the model on depth boundaries.

Main Results:

• The proposed method significantly improves quantitative and visual qualities of recovered depth maps.
• Experiments on five datasets demonstrated superior performance in depth map recovery, super-resolution, and enhancement.
• Object boundaries in recovered depth maps were accurately corrected, sharp, and well-aligned with RGB images.

Conclusions:

• The RGB-guided depth map recovery method effectively corrects boundary distortions.
• The approach offers substantial improvements over existing methods for various depth map processing tasks.
• Accurate and aligned depth boundaries are achievable, enhancing 3D scene representation.